What problem does it solve? After training a neuroimaging or tabular model, researchers need reproducible explanation artifacts—attribution maps, Grad-CAM localizations, or attention weights—without inventing explainers or weakening checkpoint security. This Skill produces target-specific explanations with numerical QA and a provenance manifest. ## Core Features & Use Cases - Integrated Gradients: Computes signed per-feature or per-voxel attributions against an explicit baseline, with trapezoidal integration and completeness checks. - Grad-CAM and attention export: Generates ReLU Grad-CAM maps from a selected 2D/3D layer, or exports real normalized attention weights (including BNT's get_attention_weights()). - Spatial NIfTI output: Writes per-subject attribution maps aligned to reference model-grid NIfTIs with validated affines, plus diagnostics.csv, qc.json, a preview figure, and run_manifest.json. - Use Case: Given a trained 3D CNN checkpoint classifying MRI scans, run the gradcam method with a chosen encoder layer to produce per-patient spatial heatmaps and a manifest binding checkpoint, preprocessing, and output hashes. ## Quick Start Run the explain script with a reviewed adapter file, checkpoint, preprocessed NPZ inputs, and metadata JSON, choosing method ig, gradcam, or attention plus a target class and output directory.
